DiffSim: Aligning Diffusion Model and Molecular Dynamics Simulation for Accurate Blind Docking
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Diffusion Model, Molecular Dynamics Simulation, Blind Docking, Drug Discovery, Molecular Conformation Generation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Predicting the ligand’s binding conformation within a target protein is a pivotal step in drug discovey. Based on prior knowledge of the binding site (protein pocket) on the target protein, biochemical researchers use molecular docking software to generate the ligand conformation within that pocket. Despite its speed, molecular docking is ill-suited for blind docking where the pocket is unknown, and the generated ligand conformation often lacks required precision. Recently, deep generative models, especially diffusion models, have been proposed for accurate blind docking. However, it is found that while deep generative models excel in locating the pocket, they still lag behind traditional methods in terms of conformation generation. Thus, bridging such gap with a hybrid approach is naturally expected to further improve the model performance. Therefore, in this study, we introduce a blind docking approach named DiffSim to seamlessly integrate the diffusion model with molecular dynamics (MD) simulation. We propose a novel loss function to align reverse diffusion sampling with MD simulation trajectories, aiming to efficiently generate ligand conformations informed by MD-modelled protein-ligand interactions at atomic resolution. Through theoretical analysis, we unveil the consistency in dynamics between diffusion models and MD simulation, demonstrating that the diffusion model is essentially a coarse-grained simulator for MD simulation. Empirical results demonstrate the effectiveness of our approach and highlight the potential of combining physics-informed MD simulation with deep learning models in drug discovery.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5830
Loading